Research questionHow can systems route each task to a suitable LLM endpoint under competing quality, cost, latency, and policy constraints?LLM endpoints can differ substantially in quality, price, latency, context support, tool use, domain expertise, and reasoning behavior. Fixed choices and manually maintained heuristics may therefore produce inconsistent trade-offs across tasks. Latest papersRecent research connected to this question, newest first.SCX Router: Streaming Zero-Shot Model Selection with a Decoder-KV Classifier and a Real-World Task OntologyThe source describes a lightweight zero-shot router that scores candidate model labels without autoregressive generation, using a 0.6B decoder with a text-only key–value cache across dialogue sessions. It also predicts task attributes and supports custom labels, while policies such as eligibility, cost, cache reuse, safety, and sovereignty remain separately specified. Evidence includes a structured task ontology, generated and open-ended tasks, and results on six LiveBench subsets, including a selected 1,000-task subset.research paper · Sep 2, 2026